vrnn_generate.py

from math import pi
import os

from keras.layers import Input
from keras.layers import Lambda
from keras.layers import LSTM
from keras.layers import Dense
from keras.layers import TimeDistributed
from keras.layers import merge
from keras.models import Model
from keras.optimizers import Adam
from keras import backend as K

from callbacks import SavePeriodicCheckpoint
from config import parse_args
from utils import audio_amplitudes_gen
from utils import samples_per_epoch
from utils import write_audio
from vrnn_model import build_vrnn


def generate(wav_dir, model, write_dir, lstm_size=1000, num_steps=40,
             z_dim=100, batch_size=32, fc_dim=400, wav_dim=200,
             num_gen=1):
    vae, decoder = build_vrnn(
        lstm_size=lstm_size, num_steps=num_steps, z_dim=z_dim,
        batch_size=batch_size, fc_dim=fc_dim, wav_dim=wav_dim, mode="generate")
    vae.load_weights(model)

    gen = audio_amplitudes_gen(
        wavdir=wav_dir, num_steps=num_steps, batch_size=batch_size,
        wav_dim=wav_dim, infinite=False)

    counter = 0
    for (x_t, y_t), true in gen:

        true_path = os.path.join(write_dir, "%d_true.wav" % counter)
        gen_dir = os.path.join(write_dir, str(counter))
        os.makedirs(gen_dir)
        for i in range(num_gen):
            pred = decoder.predict([x_t, y_t], batch_size=batch_size)

            print("Writing audio %d" % counter)
            pred_path = os.path.join(gen_dir, "%d.wav" % i)
            write_audio(pred, pred_path)
        counter += 1

if __name__ == "__main__":
    args = parse_args(mode="generate")
    generate(args.wav_dir, args.model, args.write_dir,
             z_dim=args.z_dim, lstm_size=args.lstm_size, num_steps=args.num_steps,
             batch_size=args.batch_size, num_gen=args.num_gen,
             fc_dim=args.fc_dim, wav_dim=args.wav_dim)